Uncovering nonlinear patterns in time-sensitive prehospital breathing emergencies: an exploratory machine learning study.

Background: Timely prehospital care is crucial for patients presenting with high-risk time-sensitive (HRTS) conditions. However, the interplay between response time and demographic factors in patients with breathing problems remains insufficiently understood. In this exploratory study, we applied machine learning (ML) to examine how emergency medical response time, age, and sex jointly influence the probability of encountering HRTS conditions.

Methods: A retrospective observational analysis was conducted on 132,395 prehospital missions in Stockholm (2017-2022). Multiple ML models, random forest, gradient boosting, neural networks, and logistic regression were trained to probe potential nonlinear patterns and interactions, not with the primary goal of predictive accuracy. Model performance was evaluated using sensitivity, specificity, and area under the curve (AUC) measures. However, partial dependence (PD) and individual conditional expectation (ICE) plots were the principal tools illustrating how response time, age, and sex shape HRTS likelihood.

Results: PD and ICE plots revealed that older age (> 60 years) was consistently associated with a higher probability of HRTS. Moreover, patients over 60 years displayed a complex, rising risk at prolonged response times exceeding two hours. Gradient boosting offered the best (though modest) classification metrics, with an AUC of 0.66 and an F1-score of 0.55. We emphasize that these metrics, while necessary for completeness, were secondary to our aim of characterizing nonlinear relationships.

Conclusions: Our findings underscore the exploratory value of ML in identifying subtle relationships and interactions among response time, age, and sex for time-sensitive breathing emergencies. These results highlight opportunities to refine dispatch protocols, develop age- and sex-focused screening questions, and revisit lower-priority calls after extended wait times. Future work should incorporate richer data and refine these insights for potential predictive use.

Clinical trial number: Not applicable.